Modeling and exploiting goal and action associations for recommendations

Recommender systems are used to identify those items in a large collection that are more likely to be of interest to a user. A common principle of most recommenders is that whatever happened in the past is a good indicator of the future. We offer a different perspective. Considering the fact that in real life users do their selections with certain goals in mind, we recommend items (or actions) that help users fulfilling their intended goals using their past only as a way of identifying goals of interest. We introduce a model that connects goals and actions through action sets implementing the respective goals. Such a model captures latent associations among goals and actions and allows the ranking of actions considering different user strategies such as to complete at least one goal with the minimum effort (i.e., minimum number of actions), or to open up more paths for fulfillment of more goals in the future. For each strategy we recommend an algorithm that exploits the user action and goal spaces to rank the actions in a different way. We have performed extensive experimental studies to understand how these techniques are related and compare the results against traditional recommendation methods. The experiments illustrate that it is not possible to replicate the results of our approach using existing techniques.